agricultural technology adoption and rural poverty: a study on smallholders … · 2017-07-27 ·...

Agricultural Technology Adoption and Rural Poverty:A Study on Smallholders in Amhara Regional State,

Ethiopia

Samuel Sebsibie1, Workineh Asmare2 and TessemaEndalkachew3

Abstract

Poverty has remained to be the major development and policy challenge in

Ethiopia. Studies indicate that poverty is higher in rural areas in all its

measures. Since smallholder agriculture is the mainstay of rural dwellers,

policies of the country intended to give priority to increasing the productivity of

agriculture to challenge rural poverty. Consequently, different strategies were

put in place since the 1994 reform period. Under the growth and transformation

plan, whose tenure has just come to an end, intensification (through adoption of

agricultural technologies) and structural change were sought to bring about

smallholder ‘productivity revolution’ for a transformative growth in the sectorand poverty reduction. Agricultural technology adoption is however limited in

the country with greater geographical differences. We analyze smallholders’propensity to and intensity of agricultural technology adoption in Amhara

Regional State using Double-Hurdle Model to identify the relative importance of

the factors that explain the underlying choice. A modest attempt is also made to

link technology adoption to household welfare using matching techniques of

impact evaluation. The study is based on the Ethiopian Socioeconomic Survey

(ESS, 2013/14) data of the Living Standards Measurement Study. The results

corroborate the importance of policy support schemes, input market and

physical infrastructure, poverty [capacity] to explain agricultural technology

adoption. Considerable evidence on the positive welfare impact of technology

adoption is documented which entails a tenable link between technology

1 Lecturer, School of Economics, University of Gondar, Email:[email protected] Lecturer, School of Economics, University of Gondar, Email:[email protected] Lecturer, School of Economics, University of Gondar, Email:[email protected]

Samuel, Workineh and Tessema: Agricultural technology adoption and rural poverty:...

118

adoption and poverty reduction. However, a comprehensive policy framework is

needed to tackle the capacity and physical access constraints to promote

agricultural technology adoption.

Key words: technology adoption, poverty, double hurdle, PSM.JEL code: C24, I32, Q12

1. Introduction

Food insecurity and poverty remain to be the major development and policychallenges in Ethiopia. The country is found at the tail of both the hungerindex (76th out of 79 countries) of the International Food Policy ResearchInstitute (IFPRI) and human development index (173th out of 187 countrieswith a score 0.396 in 2012) of the United Nation DevelopmentProgram(Oxfam. 2012. UNDP. 2013). An interim report by the governmentindicates that the proportion of people below the poverty line stood at29.6percent. The Disaggregated results show that those proportions in ruraland urban areas for the year 2010/11 are 30.4 percent and 25.7 percentrespectively (FDRE. 2012). Although it is recognizable that there is asubstantial decline in poverty incidence (from 38.7 percent in 2004) anddepth (from 0.083 in 2004 to 0.078 in 2010/11), the size and impact of theproblem is still considerably worrisome. It is also indicated in the interimreport that rural poverty is higher than urban poverty in all periods. More so,much of the decline in country level poverty is arguably attributed to thedecline in urban poverty albeit poverty (measured in incidence and depth)has shown a declining trend both in rural and urban areas. The reportassociated the observed improvement to the pro-poor programs implementedby the government both in rural and urban areas. However, the distributionof income among the poor showed no improvement.

On the other hand, some studies showed that rural poverty is on anincreasing trend since 2004. Studies based on the Ethiopian Rural HouseholdSurvey data, a unique longitudinal data over 15 years (1994-2009), found

Ethiopian Journal of Economics Vol. XXIII No 2, October 2014

119

that rural poverty has shown a tremendous decline up to 2004 (Dercon et al.

2007) and then started to rise (Dercon et al., 2011). Yet, movement into andout of the state of poverty – poverty dynamics – is high during all periods.These studies suggested that transition into and out of (chronic) poverty arerelated to the growth handicaps faced by the poor while the increase in ruralpoverty during 2004–2009 is closely related to local factors and highinflationary pressure. Although most of the growth stimulants (includingagricultural extension packages and infrastructure) have fundamentally thesame effect for the chronic-poor and the rest, the former face considerablysevere growth handicap compared to the rest, leaving them permanentlybehind.

The agricultural sector is believed to be the key sector both for povertyalleviation and to materialize a transformative growth in Ethiopia. This,together with its major contribution to rural livelihood in the country wherepoverty is the highest, has placed the sector at the center of development andpolicy interventions (FDRE, 2012; FDRE, 2010). More than 80percent of thepopulation lives in rural Ethiopia where its livelihood is tied to traditionalagriculture which in turn is hinged to sporadic rainfall. As a result,agricultural growth and sustainability has become a priority in policy makingin the country (Asfaw and Shiferaw, 2010; Pycroft, 2008; FDRE, 2010).Under the broader Agricultural Development Led Industrialization (ADLI)development strategy and pursuant to the prevailing structure of theeconomy, the policy environment has long been giving priority to thestructural change and productivity of the pervasive peasant agriculture.Pushing the prevailing agricultural technologies to the frontier is argued inthe ADLI framework to bring about productivity growth, inter alia,application of improved seeds, increased adoption of inputs (fertilizer.pesticide…) and expanding irrigation and infrastructure (Dercon and Zeitlin,2009). However, the rate of technology adoption and its intensity in thecountry is very low even by sub-Saharan standard. For instance, the averageadoption rate of modern fertilizer is estimated to be less than 33% of the totalcultivated land and the average level of use of modern fertilizer is only 11kgper hectare which is very low compared to 48kg per hectare in Kenya. In


120

addition, the loss of soil nutrients due to land degradation and improper useof animal dung is the highest in sub-Saharan Africa (Yesuf and Köhlin,FDRE, 2010).

Moreover, technology adoption in the country followed a clear spatialpattern as well as greater variations by crops. Amhara region is among thosewhich are characterized by low technology adoption which is mostlyattributed to fragmented land, environmental degradation, population andlivestock pressure, and relatively low productivity. North Gondar zone canbe recited in here for it is one of the least adopters in the country which iscorroborated by a recent study that argued for a limited application ofchemical fertilizer in all crops except Teff (Yu et al. 2011).

In response to the observed insufficient agricultural technology adoption andto promote intensification, several attempts have been made to identify therelative importance of factors that determine smallholders’ technologyadoption in Ethiopia and developing countries, at large. However, previousresearch generally focused on a mere characterization of farm households interms of adoption in which the socioeconomic characteristics of farmers areused to explain adoption level. The more fundamental process in which farmhouseholds make choices regarding the adoption of the available technologybased on the specific features of the technology (inter alia. suitability andlinkage with farmers’ indigenous knowledge and experience) has long beensidelined.

For a complete understanding of agricultural technology adoption and itseffectiveness, research needs to equally focus on how smallholders’ adoptionchoices and its intensity are explained by the peculiar features of thetechnology in point in relation to pre-existing farm knowledge. These factorsare of vital importance not only for adoption decision but also for theeffectiveness of the adopted technology. On the supply side, agriculturaltechnologies are often introduced in a package program although mostadopters use part of the package. And, the determinants and welfare impact


121

of such variation in adoption is not well investigated in the study area andEthiopia, in general.

The relationship between technology adoption and poverty reduction is yetambiguous. The extant literature seemed to have focused on the contributionof technology adoption in poverty alleviation in which a positive impact oftechnology adoption on household wellbeing appeared to be a generalconsensus. Among the contributions that corroborate such a relationship, theEEA/EEPRI (2006) report documented that the introduction of improvedseed and chemical fertilizer in a package program has generally a positiveimpact on productivity for although its impact on Teff production was foundambiguous (cited in (Brown and Teshome, 2008)). Most other studies alsocommend that technology adoption has a direct role on improving ruralhousehold welfare through increasing agricultural productivity (Asfaw andShiferaw, 2010).

On the other hand, some studies argue that agricultural technology adoptiondepends on poverty status of households. Although agricultural technologyadoptions have identical impact on the poor and non-poor, the poor havemore capacity constraints to adopt. Moreover, Agricultural technologyadoption is a high risk and high return choice. The farmers need to investmore to get the technology. Since the poor cannot insure their consumptionagainst shocks like crop failure, they will generally limit themselves to lowrisk low return choices (Dercon et al., 2007). Hence, poor people may notchoose to adopt agricultural intensification schemes. This is very importantyet an overlooked issue both in the literature and policy debates. Theimplication of this line of argument is that the poor smallholders refrain fromadopting productivity enhancing technologies and face a type of viciouscircle of poverty under the backdrop of no planned systematic intervention.However, this theoretical possibility shall be supported by empiricalevidence which is argued missing in this particular study.

Against this backdrop, the objectives of this paper are twofold. First, itanalyses the propensity to and intensity of agricultural technology adoption


122

by smallholders to identify the relative importance of the factors that explainthe underlying economic process. It then evaluates the impact, throughincreased productivity, of technology adoption on household welfare(poverty). With an overarching framework of the technology adoption–farmproductivity–poverty reduction nexus, the study provides empirical evidenceon the relationship between agricultural technology adoption and ruralpoverty in Amhara Region. For empirical focus, the study concentrates onadoption of chemical (inorganic) fertilizer. And analysis is made for threemajor crops.

2. Model and Estimation2.1 Understanding Smallholders’ Technology Adoption

Informed by microeconomic theories of the firm [farm], smallholders aremodeled as producing agents which decide on the use of certain technologyproducts based on its profitability. For instance, a smallholder chooses toapply chemical fertilizer on its farm if the productivity gain outweighs all thecosts associated with the use of the fertilizer. Basically, observed level ofadoption is an outcome of two distinct processes. The first stage involves theadoption decision of farmers and is commonly called participation decision.In the second stage, farmers decide on the intensity of use of the fertilizer.The standard Tobit model can be good candidates in modeling the adoptionbehavior of smallholders under no (capacity and information) constraints.However, two major drawbacks of Tobit can be considered here, especiallyfrom the perspective of our study. First, Tobit model assumes that decisionto adopt a given technology and intensity of adoption are governed byfundamentally the same stochastic process. The same vector of parametersare assumed to determine the first and the second stages of decision. Theserules out the possibility that a given variable has different marginal effectson the probability of adoption and intensity of adoption. It is also impossibleto have different vectors of parameters for the two stages of decision underTobit setting (Burke, 2009; Eakins, 2014).


123

Second, the Tobit model assumes that zero value is observed when thedependent variable is censored at zero. However, as explained in Cragg(1971), zero values may be observed due to other factors too. In the contextof our study, there are varieties of demand and supply side constraints whichcan be fairly associated with the general low rate of agricultural technologyadoption. The demand side constraints are reasonably related to incomelevels, access to credit and (most importantly. information. Farmers do nothave complete information to decide on the profitability of the technologyproduct. The supply-side constraints can be related either to low access toinformation, insufficient (incomplete) information or improper use ofinformation. Above all, access to a given farm technology is not guaranteed.So, ruling out constraints does not seem appealing.

Farm households’ choice to and level of technology adoption in the presenceof constraints gave rise to three distinct subsamples (Amare et al., 2012).The first groups of farmers are well aware, have demand and adopttechnology. The second groups do not want to adopt agriculturaltechnologies for it is not profitable at current prices. The third groups want toadopt the available agricultural technology but cannot get it due to supplyside constraints. Therefore, farmers’ choice is observed after passing twohurdles. Based on this classification, our model is developed to considerthree aspects of the fundamental choice process: decision to adopt (desireddemand for) a given technology, access to the technology and intensity ofadoption of the technology in question. It is under this backdrop that the usethe Double-hurdle model to estimate the propensity to and intensity oftechnology adoption is justified appropriate.

A parametric Double-Hurdle Model is argued appropriate in modelingempirical studies in evolving sequential decisions in two stages. It was firstproposed by Cragg (1971) and used in variety of empirical literatureincluding health economics(Jones, 1989; Labeaga, 1999; Tauras, 2005),estimating expenditure(Yen and Jensen, 1996; Lin and Milon, 1993), laboreconomics. valuation studies (Saz-Salazar and Rausell-Koster, 2008; Oseni,2015) and technology adoption(Islam et al., 2015; Akpan et al., 2012;


124

Hazarika et al., 2015; Asfaw et al., 2011; Gebremichael and Gebremedhin,2014). The Double-Hurdle model is a generalization of the Tobit modeldesigned to deal with survey data in which the decisions can be modeled asdependent. independent or sequential to each other(Gao et al., 1995).

Following (Amare et al., 2012; Asfaw et al., 2011), we specify the doublehurdle model as in what follows. Suppose now that for any individual farmhousehold i. the desired demand for fertilizer is given by:

1...........................'*iiXD (1)

Where Xi is a vector of determinants of demand. is a vector of parameters.

is Gauss-Markov’s error term, and ‘D*’ is latent desired demand. Inaddition, assume that the farm household’s access to fertilizer is given by

)2(..............................'*iiZA

(2)

where A* is latent variable denoting farm household’s possibility of accessto fertilizer supply; Z is vector of determinants of access to fertilizer; is

vector of parameters and is Gauss-Markov’s error term. These twoequations are assumed to be independent of each other and divide the totalsample in to three sub-samples.

1. Those households who adopt fertilizer (D*>0 and A*>0).2. Those households who do not want fertilizer regardless of access for it

(D*<0).3. Those households who have positive desired demand to fertilizer but do

not adopt due to lack of access (D*>0 and A*<0).

Hence, adoption of fertilizer is observed after it passes two thresholds:positive desired demand and access thresholds. Yet, an important decision offarm households is intensity of adoption, i.e.


125

*YYi if 0,0 ii AD (3)

otherwiseYi ,0

iiHY '*.

Where, H is a vector of variables; is vector of parameters and is Gauss-Markov’s error term (Beshir et al., 2012). On the basis of this setting, thelikelihood function for the observed demand is given by:

)4....(..........'1*'ln'1ln*'lnln12 1311

G Gi

ii

G

iii ZXXXDZL

(4)

Where, and are (resp.) probability density function (pdf) and cumulativedistribution function (cdf) of standard normal variable (Asfaw et al., 2011).

Double hurdle models with continuous response variable in the second stageare mostly estimated by specifying binary choice for the first hurdle andordinary least square (OLS) regression for the second hurdle, assuming thatthe distribution of the error terms is bivariate normal. However, in our case,the distribution of the response variable for the second hurdle was highlyskewed. Thus, normality assumption doesn’t seem to hold. The first naturalchoice would be logarithmic transformation of the response variable.However, such transformation may lead to bias in estimating elasticities asdiscussed in Tauras (2005). More so, retransformation is not easy in the caseof heteroscedastic errors (Ornelas-Almaraz. 2012. Tauras. 2005).

Whenever such distributional issues arise, GLM with log-link relationshipand appropriate distribution family is preferred (Tauras. 2005. Ornelas-Almaraz. 2012. Manning and Mullahy. 2001). GLM provides a flexibleoption to relax the normality assumption with no need for retransformation

as predictions are based on raw scale (Salmon and Tanguy, 2015; Jones,

2010;. Tauras, 2005). However, it is worth mentioning that GLM estimators

may be less precise especially for data with heavier tails in log scale (Baser,


126

2007; Manning et al., 2002). Our study uses GLM with log link and

Gaussian distribution in the second hurdle.

2.2. Welfare/Poverty Impact of Agricultural Technology Adoption

Evaluation of the welfare [poverty] impact for agricultural technologyadoption has something to do with the determination of ex-post measuredoutcomes of welfare indicators for technology adopting smallholders incomparison to the counterfactual [outcomes of welfare indicators had thesmallholders not adopted the technology product] (Heckman and Vytlacil,2005). In effect, technology adoption is considered as a treatment[intervention] which is however subject to non-random assignment tosmallholders and self-selection.

More so, the outcome variables of welfare indicators are not observable inboth treated and untreated states. This necessitate the statistical constructionof a suitable counterfactual in the untreated state conditional on receivingtreatment (Diaz and Handa, 2004). For our study of exploring the welfareimpact of agricultural technology adoption is based on observational datathan in an experimental setting. the most widely employed technique ofPropensity Score Matching is used to settle the counterfactual problem

(Austin, 2011; Steiner and Cook 2013). Apparently, the validity of matching

technique relies on certain assumptions. The first basic (identification)assumption is the conditional independence assumption. The assumptionstates that outcomes in the untreated state are independent of programparticipation conditional on a particular set of observable characteristics

(Diaz and Handa, 2004; Khandker et al., 2010). Suppose X denotes a set of

observable characteristics, and T is a dummy variable for treatment. If theparameter of interest is average treatment effect (ATE), the identificationassumption is given by:

5.................................|, XPTYY ct (5)


127

Where the symbol indicates independence and XP is the propensity

score. tY and cY indicate the outcome of interest for treated and untreatedgroups respectively. For estimation of the treatment effect on the treated(TOT), the assumption can be relaxed to

6.................................| XPTY c (6)

The second assumption of PSM is the common support condition whichrequires that the treatment observations have comparison observationsnearby in the distribution of the propensity scores. It is given by.

(7)

For estimation of TOT. the common support condition can be relaxed as

(8)

Imposing conditional mean independence assumption, our parameter ofinterest, average treatment effect on the treated (ATT), is evaluated as:

9........................,1|,1| 01 XPTYEXPTYExATT tt (9)

The second term in equation (9) is the average welfare outcome of treatedindividuals had they not been treated. However, this is not observable incross-sectional studies like ours. Instead, corresponding outcomes foruntreated observation is estimated as

10........................,0|,1| XPTYEXPTYExATT ct (10)

The difference between (9) and (10) is attributed to selection bias. In ourstudy, balancing scores are estimated from logit model and the common

7.................................1|10 XTP

8.................................1|1 XTP


128

support condition was imposed. Matching estimators, based on (Diaz andHanda, 2004; Khandker et al., 2010), have the general form as:

11..........................................................,1

Ti Cj

cj

tTi

t

YjiYN

ATT (11)

Where N is the number of participants and; ji, represents a weighting

function that depends on the specific matching estimator. Thus, the choice ofthe matching technique is crucial. While it is possible to select a matchingtechnique based on its performance in minimizing bias, we prefer to usethree commonly used matching criteria that, we believe, complement eachother. Nearest Neighbor Matching (NNM), Radius Matching (RM) andKernel Matching (KM) are used to make sure that our results are robust. Asdiscussed in Khandker et al. (2010). NNM matches each treatment unit to acomparison unit with the nearest propensity score. However, the differencein propensity score between the closest treatment and control units may stillbe high. This may be avoided by specifying the maximum propensity scoredistance (caliper) which justifies the use of Radius Method. On the otherhand, Kernel Method is a nonparametric matching estimator which usesweighted average of all nonparticipants to construct the counterfactual matchfor each participant. Major advantage of this method over the other two is

that more information is used (Khandker et al., 2010, Caliendo and

Kopeinig, 2005). Estimation was done using STATA 12 software.

The Foster-Greer-Thorbecke (FGT) index (Foster et al., 1983), which isdefined as in what follows, is used to measure the poverty status ofsmallholders.

12.......................1

1

aq

ia z

yiz

NP

(12)

Where, ‘z’ is the poverty line; y is real Percapita consumption expenditureand ‘a’ is the poverty aversion parameter. The poverty incidence, depth andseverity are measured by changing the values of ‘a’ in the formula. The


129

2010/11 national poverty line estimate, adjusted for price changes is used tocompute the FGT indices (FDRE, 2012).

3. Data and Descriptive Statistics

The study is based on data from the Ethiopian Socioeconomic Survey (ESS.2013/14). The ESS is a collaborative project between the Central StatisticsAgency of Ethiopia (CSA) and the World Bank Living StandardsMeasurement Study- Integrated Surveys of Agriculture (LSMS-ISA) team. Itprimarily targets on developing and implementing a multi-topic survey thatmeets Ethiopia’s data demand and gaps and is believed to be of high quality,accessible to the public and aligned with the National Strategy for theDevelopment of Statistics (NSDS).

The survey covers rural areas, large towns and small towns in all regionsexcept some exceptional zones of Afar and Somali region. Sample unitswere selected using stratified two stage sampling procedure and the samplingframe for rural areas was based on 2013/14 Agricultural sample survey ofthe CSA. A total of 5.262 sample households were selected for the wholecountry of which 20.5 percent (1080) are drawn from Amhara Region. Thedata is argued to be representative for regional estimation in the mostpopulous regions (Amhara, Oromiya, SNNP and Tigray) (CSA, 2011/12).As such, the survey covered 61 rural, 15 medium and large urban and 10small urban enumeration areas in Amhara regional state. The data for allrural households in Amhara region is used in the study. And. the analyses onthe pattern of technology adoption and its implied impact on rural povertyare made based on three crops, namely Teff, Wheat and Maize.

3.1 Distribution of Plot Size, Technology Adoption and Intensity

A simple exploration of the data offer important insights on adoption ofagricultural technologies and intensity of use. It is apparent from Table 1 thatabout 65% of the samples adopt chemical fertilizer with a small variationacross crop type. This is consistent with previous results in other parts of the


130

country and sub-saran Africa (Terefe et al., 2013; Endale, 2011). On the

other hand, the adoption of improved seed appeared small when compared tochemical fertilizer.

Table 1: Distribution of plot size and technology adoption by cropcategory (%)

CropCode

Improved Seed(percentage of

plots)

Fertilizer Use (percentage of plots)ChemicalFertilizer

Manure Compost

MAIZE 34.75 66.15 46.79 18.0

TEFF 3.68 66.58 9.50 8.0

WHEAT 9.80 63.91 17.40 15.86

Total 16.74 65.77 25.31 13.62

On average, only 16.74% of the plots use improved seeds with a largevariation between crop types. The adoption of improved seeds is larger formaize followed by wheat. The application of organic fertilizers (manure andcompost), as indicators of pre-existing knowledge and technological practices,is way below the application of chemical fertilizers with considerable variationacross crops. The data entails a significant variation in the intensity ofimproved seeds and fertilizer application, as well, from Table 2, the averageuse of urea and dap in kilogram per hectare of fertilized (for the three crops)land is 74.28 and 89.91 respectively. And. the intensity of chemical fertilizeruse is generally larger for maize followed by wheat.

Table 2: Distribution Fertilizer use by crop category (kg/ha)

Fertilizer TypeCrop Type

Maize Teff Wheat Total

Urea 99.15 52.85 68.89 74.28

Dap 115.62 66.36 86.68 89.91

Total (Urea + Dap) 215.16 119.35 155.21 164.29


131

4.2 Average Productivity and Technology Adoption

An important issue worth exploring is the relationship between landproductivity (yield) and application of chemical fertilizer in comparison topre-existing technological practices. An insightful result on average yieldunder different technology regimes is presented in Table 3. The averageyield is about 1461kg/ha with significant variation across crop types. Maizegives highest yield per hectare followed by wheat. Another important featurein the table is the impact of fertilizer use on productivity. There is a negativedifferential in productivity between adopters and non-adopters of bothorganic and inorganic fertilizer.

An exception in this respect is Maize for which all types of fertilizers exceptcompost have negative effect. This is not surprising for many variableswhich affect the relationship are not controlled. A typical example is cropdamage. Crop damage is reported on about 22% of the total plots taken inthe sample, and maize plots experience the largest damage both in terms ofthe number of plots with reported damage (30%) and the perceived share ofdamage from the total crop in the plots (40%). The three major causes of thedamage have been insects (21%), shortage of rain (20%) and hail (13%) [seeAppendix 2]. More so, Maize is grown in arid and semi-arid parts of theregion that makes the crop yield vulnerable to weather related and othershocks, as a result of which, the productivity impact of fertilizer is sounpredictable.

Table 3: Average Yield of Crops (in kg/ha) under Different FertilizerRegimes

CropCode

Fertilizer Regimes

TotalChemical Fertilizer Manure use CompostNo Yes Diff No Yes Diff No Yes diff

Maize 1990.9 1887.6 103.3 2163.2 1673.7 489.7 1842.7 2166.8 -324.1 1921.7

Teff 971.8 1153.8 -181.9 1073.2 1081.2 -8 1215.5 1337.6 -122.1 1092.8

Wheat 1161.0 1423.9 -262.9 1256.5 1497.9 -241.4 1454.7 1457.7 -3 1329.8

Total 1393.6 1495.6 -101.9 1585.0 1568.3 16.7 1529.1 1780.3 -179.2 1461.1


132

Generally, further exploration of the summary statistics for the study variableand its covariates offered appealing results in view of the processunderpinning the observed technology adoption and intensity of use [SeeAppendix 1]. A simple mean comparison test between adopters and non-adopters of fertilizer indicates that a set of household characteristics,including education and livestock wealth (in TLU), are significantly largerfor adopters. The mean distance from markets, all weather road and urbancenters is significantly larger for non-adopters indicating that non-adoptershave lesser access to market and technology information which in turnresults in slow diffusion of farm technology as well as high transportationcost. There is also a significant variation in the characteristics of plots withand without chemical fertilizer use. More so, adopters have significantlyhigher mean values in terms of microclimate indicators like potentialwetness index and elevation, and lower mean values in terms of plot slopeand distance from homestead.

Other factors such as extension contact, advisory service, the use of manureand compost and credit have also statistically significant association withadoption of fertilizer. The first two are related with farmers’ access toinformation on fertilizer and its profitability while access to credit indicatesfarmers’ ability to finance their purchase of modern technology under cashconstraints. The institutional support system has long been a major factor formodern technology adoption and productivity of smallholders. However,such support has remained low. The use of manure and compost (organicfertilizer) has negative association with adoption of chemical fertilizer with apossible explanation that the pre-existing technology practices are preferredsubstitutes to inorganic chemical fertilizers.

4. Estimation Results and Discussion4.1 The Probability and Intensity of Agricultural Technology

Adoption

The estimation of the probability equation for farm household’s technologyadoption, application of fertilizer on the farm, and the level of technology


133

adoption is performed under the double-hurdle model assumptions. Theestimated model relates smallholder farm households’ adoption decision andthe intensity of technology use to a long list of socio-economic factors inAmhara Region. For the interest of interpretation and discussion, we groupedsuch factors into household characteristics, plot characteristics and micro-climate factors, institutional factors and policy support. The double-hurdlemodel estimation result is presented and discussed in this section.

4.1.1 Explaining Technology Adoption in Amhara Region

Understanding technology adoption goes beyond the simple characterizationof factors as determinants of the technology use to evaluating the relativeimportance of competing theoretical explanations. Evidence from theestimated relationship on the widely discussed socio-economic factors andtechnological practices and smallholders’ technology adoption entails more onthis (Table 4). From among the household characteristics, formal educationand sex have the expected sign but in significant. Off farm employment of thehousehold head significantly increase the probability of adoption, suggestingthat smallholders diversifying into the off-farm economy and creditconstrained finance chemical fertilizer through off farm earnings. About two-third of the participants in off farm activities have no credit access and thisstrengthens the argument. Similarly, land area has significant positive impacton propensity to adopt inorganic fertilizer. Land and off farm employment arepoverty related variables. Especially land is the major component of wealth ofrural households. A simple mean comparison test indicates that the landholding of the poor4 is significantly lower than the non-poor [See Appendix 5].Our estimation result shows that the probability of adoption increases with anincrease in size of farm land owned.

4Poverty is calculated based on the national poverty line after the households’ annualPercapita expenditures are adjusted for inflation. See FDRE 2012, Ethiopia’sProgress Towards Eradicating Poverty: An Interim Report on Poverty AnalysisStudy In: Directorate, D.P.A.R. (ed.). Addis Ababa: Ministry of Finance andEconomic Development.


134

Agricultural market (input and output) and physical infrastructure constitutesanother important set of factors that influence smallholder’s technologyadoption. Accordingly, distance from the nearest all weather road, nearestmarket and zonal town have also negative and significant impact on theprobability of adoption.

The marginal effects (for the total sample) indicate that at average anincrease in distance from the nearest all weather road and nearest market by1 kilometer results in a decrease in the likelihood of chemical fertilizeradoption by 0.013 and 0.0016, respectively [see Appendix 4]. Therefore, incomparison road infrastructure has the strongest impact. Lack of roadinfrastructure is one of the major bottle necks which affect farmers’ adoptiondecision in different ways. First, lower road infrastructure increases thetransportation cost for chemical fertilizers and thus creates a large differencebetween the actual price and the price farmers’ face in the input market. Italso reduces competition between input suppliers resulting in little choiceavailable for farmers. Second, poor infrastructure reduces the profitability ofa given technology. In the output market, it increases the difference betweenfarm-gate price and the actual market price and creates geographic barrierrestricting local demand to depend only on local supply. If there isinformation asymmetry, it is possible for middlemen with better informationto expropriate information rent from farmers. Poor road infrastructure is alsoassociated with slow diffusion of agricultural technology. Distance from thenearest market and proximity to urban center also have similar effect.

Institutional support factors are considered equally important to understandsmallholders’ technology adoption. And, this is supported by the empiricalevidence. Credit and extension appeared important determinants of adoption.Around 40% non-adopters cite lack of financial capital as their major reasonfor not adopting chemical fertilizer. The two major reasons for inability toaccess credit, as reported by respondents in the study area, are inability topay previous loans (48.05%) and inadequate service (13%), implyingpervasive credit constraint for smallholders [See Appendix 6].


135

Table 4: Estimation Results of Smallholders’ Propensity to TechnologyAdoption

VariablesProbit model. Dependent variable: adoption

Maize Teff Wheat Total sampleHousehold characteristics

Sex -0.0233(0.213)

-0.303(0.287)

-0.0743(0.265)

-0.0858(0.131)

Age -0.0124*(0.00527)

-0.00630(0.00597)

-0.0106(0.00570)

-0.00901**(0.00295)

Education -0.0150(0.0471)

0.0512(0.0538)

0.0592(0.0452)

0.0376(0.0252)

Land area (hectare) 2.504***(0.685)

1.352(0.739)

-0.0397(0.420)

0.796**(0.276)

Off farm employment 0.155(0.316)

0.331(0.400)

1.183***(0.357)

0.647***(0.185)

Input market and infrastructureDistance from road all weather road -0.0193*

(0.00836)0.00509

(0.00620)-0.0205**(0.00682)

-0.00699*(0.00324)

Distance from market 0.00186(0.00384)

-0.0153***(0.00377)

-0.00878**(0.00331)

-0.00906***(0.00172)

Distance from zonal town -0.00687***(0.00157)

0.00761***(0.00184)

-0.00410**(0.00144)

-0.000880(0.000751)

Institutional supportExtension 1.743***

(0.184)2.460***

(0.250)2.103***

(0.187)1.987***(0.0997)

Credit 0.212(0.165)

0.526**(0.202)

0.487**(0.186)

0.341***(0.0935)

Use of organic inputsManure use -1.077***

(0.168)-0.651*(0.254)

-0.858***(0.260)

-0.981***(0.106)

Compost use -0.148(0.198)

-1.276***(0.293)

-0.622*(0.277)

-0.572***(0.123)

Plot characteristics

Plot distance from home -0.0541(0.0610)

-0.0880(0.0714)

-0.0252(0.0438)

-0.0264(0.0285)

Soil quality (poor=1) 0.165(0.123)

-0.376**(0.130)

-0.0816(0.117)

-0.149*(0.0620)

Plot slope -0.0464***(0.0128)

-0.0284***(0.00837)

-0.0115(0.00774)

-0.0285***(0.00470)

Plot potential wetness index 0.0531(0.0597)

-0.00738(0.0621)

-0.0586(0.0310)

-0.0264(0.0209)

Altitude 0.00123***(0.000342)

0.000800**(0.000269)

0.000272(0.000238)

0.000596***(0.000112)

Agroecology -0.914**(0.333)

-0.278(0.257)

-0.297(0.200)

-0.361**(0.131)

_cons -1.503(1.111)

-1.677(1.259)

2.091*(0.970)

0.271(0.463)


136

Akin, plots under extension are more likely to adopt chemical fertilizer in theregion. Agricultural extension is the major instrument for dissemination ofoutputs of agricultural research. In Ethiopia, though agricultural extensionhas a long history, the dissemination of technology has been less thanexpected. Agricultural extension affects technology adoption decision inmany ways. First, extension workers give training and advisory service tofarmers which increase human capital and information access. Second,agricultural extension is mostly coupled with input distribution and farmcredit. Third, it is the major channel through which agricultural research anddevelopment outputs are transferred to smallholders.

Plot characteristics and microclimate variables are another set of factorsconsidered in explaining smallholders’ technology adoption, mainly toaccount for varying plot quality and unobserved differences between agro-ecological zones. The three crops are grown in two major agro-ecologicalzones of the region: Semi-arid and sub-humid. The result shows that thelikelihood of adoption in semi-arid areas is higher and significant for maize.On the other hand, plot slope and altitude have negative and significanteffect on adoption. The probability of adoption decreases with deterioratingsoil fertility indicating that framers are less likely to adopt inorganicfertilizer on poor quality plots. Another important implication of the result isthe relationship between adoption of inorganic and organic fertilizers [SeeAppendix 7]. The use of manure and compost has strong negative impact onthe adoption of inorganic fertilizer corroborating the result of possiblesubstitutability in the descriptive analysis.

4.1.2. Explain the Intensity of Technology Adoption in Amhara Region

Study of agricultural technology in relation to smallholder productivity andimplied impacts on poverty reduction will be fairly complete when analysisof farm households’ propensity to technology adoption is substantiated toanalysis of the intensity of technology use. In line with this strand ofthinking, intensity equation is estimated for fertilizer use, the result of whichis presented in Table 5. Household characteristics such as age and livestock


137

ownership have negative and significant effect while education has positiveand significant effect. The negative effect of livestock ownership onintensity of use may be due to that livestock ownership increases householdsaccess to manure which they use as a substitute for chemical fertilizer. Giventhe transportation cost, poor infrastructure and other constraints, farmersmay prefer to use manure though livestock ownership also relaxes their cash

constraint. Similar results were found by other studies(Hailu et al., 2014;Kassie et al., 2009). The effect of off farm employment is mixed.

Off-farm employment has positive and significant effect for Teff and wheatbut negative effect for maize. On the other hand, access to all weather roadsand other market related variable has mixed effects by crop type but are notsignificant for the total sample. Access to extension has a positive andsignificant effect on intensity of adoption which, as discussed in the adoptiondecision part above, may be due to that access to extension is the major waythrough which farmers get technology information and other servicesimportant for. Plot size has a negative and significant effect on intensity offertilizer use. With regard to plot characteristics that are related withmicroclimate, the results indicate positive and significant effects of potentialwetness index of soil, altitude and agro-ecology.


138

Table 5: Estimation Results of Smallholders’ Intensity of TechnologyAdoption

GLM

Dependent Variable: Intensity of Adoption (amount of fertilizerin kilogram per hectare of land under chemical fertilizer).logarithm link and Gaussian distribution

Maize Teff Wheat Total sampleHousehold characteristicssex 0.302*

(0.130)-0.0145(0.199)

-0.421(0.526)

0.472*(0.200)

age -0.00773**(0.00290)

-0.0156**(0.00510)

-0.111***(0.0254)

-0.0122**(0.00429)

Education -0.0126(0.0235)

0.0642**(0.0224)

0.288***(0.0490)

0.164***(0.0138)

Household size 0.00203(0.0258)

0.0765*(0.0304)

0.444**(0.136)

-0.0964**(0.0309)

Off farm employment -0.217(0.145)

0.406*(0.198)

1.134**(0.440)

-0.836**(0.318)

Livestock ownership (tlu) -0.0390*(0.0174)

-0.00209(0.0243)

-0.850***(0.183)

-0.153***(0.0312)

Input market and infrastructure relatedDistance from road allweather road

-0.0111(0.00664)

-0.00132(0.00469)

-0.179***(0.0477)

-0.00914(0.00468)

Distance from market -0.00384(0.00222)

-0.00439(0.00262)

0.0273***(0.00757)

0.00215(0.00175)

Distance from zonal town -0.000834(0.00120)

0.00348*(0.00156)

0.0115**(0.00382)

0.00206(0.00126)

Institutional supportExtension 0.378*

(0.156)0.823***

(0.196)1.203**(0.401)

0.848***(0.174)

credit -0.0446(0.0877)

0.144(0.134)

-0.454(0.296)

-0.00174(0.106)

Plot characteristicsAgro- ecology 0.555

(0.303)0.370

(0.248)3.580***

(0.905)0.490*(0.249)

Plot size -0.853**(0.325)

-2.812***(0.596)

-24.78***(3.043)

-7.780***(0.836)

Plot slop -0.00904(0.0100)

-0.0285**(0.00944)

0.00957(0.0144)

0.00619(0.00439)

Plot wetness index 0.00587(0.0277)

0.00964(0.0335)

0.147*(0.0681)

0.0876**(0.0288)

Altitude 0.000456*(0.000208)

0.0000351(0.000174)

0.00184***(0.000315)

0.000712***(0.000128)

Constant 4.583***(0.678)

4.772***(0.814)

0.301(1.603)

2.770***(0.606)

N 661 415 653 1729Standard errors in parentheses * p<0.05. ** p<0.01. *** p<0.001


139

4.2 Welfare/Poverty Reduction Impact of Technology Adoption

Rural poverty reduction and food security has long remained to be thepriority of poor agricultural economies, of which Ethiopia is an excellentcase in point, in introducing agricultural innovations. Evidence-based policychoices can be motivated in this particular area with the analysis ofwelfare/poverty impacts, through productivity growth, of technologyadoption. This section presents a modest attempt in this direction in whichthe impact of fertilizer use on household welfare is analyzed and discussionsof the results ensue. To fix ideas, we use (1) propensity score matchingtechnique and (2) simple comparison of poverty incidence, depth andseverity between adopters and non-adopters of fertilizer. Logit model is usedto compute the propensity scores in the first method and FGT curves are fitletting the poverty line to vary in the interval [0.7562] to ensure robustcomparison in the later5.

4.2.1 Comparison of Poverty Indices by Technology Adoption

The poverty rate for the region is estimated at 28.8%. only slightly less thanthe 2010/11 estimated poverty rate for the rural Amhara which wasestimated at 30.7%(FDRE. 2012). This indicates that no significantimprovement has been made in reducing rural poverty in the region for thelast three years until 2013/14. On the other hand, it is evident from Table 6that adopters have lower outcomes in terms of headcount, depth and severityof poverty. The poverty headcount for adopters is 20.8% while the same fornon-adopters is 38.6%. The poverty curves are also fitted to make thecomparison independent of the choice of poverty line [See Appendix 8]. Thepoverty incidence, depth and severity curves show a clear dominance fornon-adopters in that adopters have lower incidence, depth and severity at allpossible poverty lines in the range [0. 7526].

5The upper limit is selected to be twice the poverty line considered in our analysis.


140

Table 6: Poverty Headcount. Depth and Severity of for Adopters andNon-Adopters

Index Adopter Non-adopter Total

Head count 6 0.208281(0.023613)

0.386223(0.031795)

0.287942(0.019529)

Depth0.096318

(0.012774)0.174695

(0.017320)0.131406

(0.010550)

Severity0.057345

(0.008879)0.104598

(0.012231)0.078499

(0.007383)

This gives a clue on the positive welfare (poverty reduction) impact oftechnology adoption. However, it is worth mentioning that such comparisondoesn’t account for problems of confounding which necessitates‘ impactevaluation proper’ for it will single out the impact of adoption. Simplecomparison of poverty by adoption status may correctly reflect the effect oftechnology adoption as poverty its self may the factor for not adoptingagricultural technology. Therefore, we use propensity score matchingtechnique to evaluate the impact of technology adoption under statisticallycontrolled environment.

4.2.2 Estimation of Average Treatment Effects for Technology Adoption

The ATT impact of agricultural technology adoption on poverty reduction(welfare) was estimated using Kernel Matching. Nearest NeighborhoodMatching and Radius Matching Methods. Estimated average treatmenteffects from the three matching methods are presented for purposes ofcomparison and robustness check [Table 7]. For the interest of clarity, thecovariate balancing test procedure is performed to check whether thedistributions of relevant covariates of adoption are balanced before and aftermatching, once the assumptions of the model are satisfied [Table 8].

6Standard errors in parenthesis


141

Table 7: Summary of Covariate Balancing TestA

lgor

ithm

Psedo R2 LR 2 (p-values) Mean bias

Tot

al %

|bia

s|re

duct

ion

Beforematching

Aftermatching

Beforematching

Aftermatching

Beforematching

Aftermatching

NNM 0.160 0.013117.88(0.000)

11.11(0.196)

30.9 6.6 78.6

Kernel 0.160 0.006117.88(0.000)

5.33(0.722)

30.9 6.0 80.58

Radius 0.160 0.005117.88(0.000)

4.66(0.793)

30.9 5.5 82.2

The likelihood ratio test for joint significance of the covariates is stronglysignificant before balancing and insignificant after balancing for allmatching algorithms. In addition, the pseudo R2 is very small after balancingindicating that the model balances the covariates between adopters and non-adopters. On the other hand, the bias minimizing matching algorithm isfound to be the radius matching. Caliper size is another important point to benoted. In our case, (for NNM and Radius matching algorithms) caliper sizeof 0.056, determined based on the recommended way as 1/4th of the standarddeviation of the propensity score. is used (Rosenbaum and Rubin, 1985).

It is evident from the result in Table 8 that technology adoption significantlyimproves household welfare (reduce poverty) as measured by Percapitaconsumption expenditure. Per capita consumption expenditure has increasedby about 16.9% and 23.9% with the KM and RM techniques, respectively.The estimated positive impact of agricultural technologies supports thetheoretical explanations argued.

Table 8: Estimation of Average Treatment Effect on the Treated

Matching MethodN

ATT Std error t-valueTreated Control

Kernel matching (KM) 306 232 0.169 0.081 2.098

Nearest neighbour (NNM) 306 105 0.154 0.109 1.407

Radius matching (RM) 306 232 0.239 0.078 3.084


142

Sensitivity of the average treatment effect in the presence of unobservedheterogeneity between the treatment and control groups is tested usingRosenbaum bounds test [See Appendix 9]. The results show that the upperand lower bound estimates of significance levels for changes in gammavalues at 0.05 intervals between 1 and 2 are zero, ensuring the robustness ofthe estimated ATT.

5. Conclusion and Policy Implication

The governments in developing countries, where structural transformation isyet to take place, have long been through agriculture led growth to addressdevelopment challenges of poverty and food security. Agriculturalinnovation through enhanced technological capabilities has is emerging as aconsensus paradigm in which systems of agricultural research anddevelopment, technology transfer and adoption are seriously considered togo about. Under the growth and transformation plan recently concluded.intensification (through adoption of agricultural technologies) and structuralchange were sought to bring about smallholder ‘productivity revolution’ fora transformative growth in the sector and poverty reduction. Agriculturaltechnology adoption is however limited in the country with greatergeographical differences. In view of this, we analyze smallholders’propensity to and intensity of agricultural technology adoption in AmharaRegional State using Double-Hurdle Model. The Ethiopian RuralSocioeconomic Survey (ESS. 2013/14) data of the Living StandardsMeasurement Study is used to estimate key relationships. The resultscorroborate the importance of policy support schemes. input market andphysical infrastructure, poverty [capacity] to explain agricultural technologyadoption. Considerable evidence on the positive welfare impact oftechnology adoption is also documented which entails a tenable link betweentechnology adoption and poverty reduction policy endeavors. On the basis ofthis findings, the study urges for comprehensive policy frameworks to tacklethe capacity and physical access factors which deter farmers from adoptingagricultural technology..


143

References

Akpan, S. B., Nkanta, V. S. & Essien, U. A. (2012). A Double-Hurdle Model ofFertilizer Adoption and Optimum use among Farmers in Southern Nigeria,Tropicultura, 30, 249-253.

Amare, M., Asfaw, S. & Shiferaw, B. (2012). Welfare Impacts of Maize–pigeonpeaIntensification in Tanzania. Agricultural Economics, 1-17.

______. (2010). Agricultural Technology Adoption and Rural Poverty: Applicationof an Endogenous Switching Regression for Selected East AfricanCountries. Poster Presented at the Joint 3rd African Association ofAgricultural Economists (AAAE) and 48th Agricultural EconomistsAssociation of South Africa (AEASA) Conference, Cape Town. SouthAfrica

Asfaw, S., Shiferaw. B., Simtowe, F. & Haile, M. G. (2011). AgriculturalTechnology Adoption, Seed Access Constraints and Commercialization inEthiopia. Journal of Development and Agricultural Economics. 3, 436-447.

Austin, P. C. (2011). An Introduction to Propensity Score Methods for Reducing theEffects of Confounding in Observational Studies. University of Toronto:Institute for Clinical Evaluative Sciences Department of HealthManagement, Policy and Evaluation.

Baser, O. (2007). Modeling Transformed Health Care Cost with UnknownHeteroskedasticity. Applied Economic Research 1-6.

Beshir, H., Emana, B., Kassa, B. & Haji, J. (2012). Determinants of ChemicalFertilizer Technology Adoption in North Eastern Highlands of Ethiopia:The Double Hurdle Approach. Journal of Research in Economics and

International Finance (JREIF). 1, 39-49.Brown, T. & Teshome, A. (2008). Implementing Policies for Chronic Poverty in

Ethiopia. Background Paper for the Chronic Poverty Report. ChronicPoverty Research Center.

Burke, W. J. (2009). Fitting and Interpreting Cragg’s Tobit Alternative using Stata.The Stata Journal. 9, 584-592.

Caliendo, M. & Kopeinig, S. (2005). Some Practical Guidance for theImplementation of Propensity Score Matching. IZA Discussion Paper

Cragg, J. G. (1971). Some Statistical Models for Limited Dependent Variables withApplication to the Demand for Durable Goods. Econometrica 39, 829-844.


144

Dercon, S., Hoddinott, J. & Woldehanna, T. (2007). Growth and Poverty in RuralEthiopia: Evidence from 15 Communities 1994-2004. Background Paperfor the Chronic Poverty Report Chronic Poverty Research Center.

_______. (2011). Growth and Chronic Poverty: Evidence from Rural Communitiesin Ethiopia. CSAE Working Paper WPS. 18.

Dercon, S. & Zeitlin, A. (2009). Rethinking Agriculture and Growth in Ethiopia: AConceptual Discussion. Paper Prepared as part of a study on Agricultureand Growth in Ethiopia.

Diaz, J. J. & Handa, S. (2004). An Assessment of Propensity Score Matching as aNon Experimental Impact Estimator: Evidence from a Mexican PovertyProgram.

Eakins, J. (2014). An Application of the Double Hurdle Model to Petrol and DieselHousehold Expenditures in Ireland. Survey Energy Economics DiscussionPaper Series.

Endale, K. (2011). Fertilizer Consumption and Agricultural Productivity in Ethiopia.EDRI Working Paper 003.

FDRE. (2010). Ethiopia’s Agricultural Sector Policy and Investment Framework(PIF) 2010-2020. Ministry of Agriculture and Rural Development.

______. (2012). Ethiopia’s Progress Towards Eradicating Poverty: An InterimReport on Poverty Analysis Study: Directorate, D.P.A.R. (ed.), AddisAbaba: Ministry of Finance and Economic Development.

Foster, J., Greer, J. & Thorbecke, E. (1983). Notes and Comments: a Class ofDecomposable Poverty Measures. Econometrica. 52, 761-766.

Gao, X. M., Wailes, E. J. & Cramer, G. L. (1995). Double-hurdle Model withBivariate Normal Errors: An Application to U.S. Rice Demand. Journal of

Agriculture and Applied Economics. 27, 363-376.Gebremichael, B. & Gebremedhin, B. (2014). Adoption of Improved Box Hive

Technology: Analysis of Smallholder Farmers in Northern Ethiopia.International Journal of Agricultural Economics and Extension 2, 077-082.

Hailu, B. K., Abrha, B. K. & Weldegiorgis, K. A. (2014). Adoption and Impact ofAgricultural Technologies on Farm Income: Evidence from SouthernTigray. International Journal of Foodand Agricultural Economics. 2, 91-106.

Hazarika, B., Bezbaruah, M. P. & Goswami, K. (2015). Adoption of ModernWeaving Technology in the Handloom Micro-Enterprises In Assam: ADouble Hurdle Approach. Technological Forecasting & Social Change

Heckman, J. J. & Vytlacil, E. (2005). Structural Equations, Treatment Effects andEconometric Policy Evaluation 1. Econometrica. 73, 669-738.


145

Islam, A. H. M. S., Barman, B. K. & Murshed-E-Jahan, K. (2015). Adoption andImpact of Integrated Rice–Fish Farming System in Bangladesh.Aquaculture 447, 76-85.

Jones, A. M. (1989). A Double-Hurdle Model of Cigarette Consumption. Journal of

Applied Econometrics. 4, 23-39.Jones, A. M. (2010). Models For Health Care. HEDG Working Paper 10/01.Kassie, M., Yesuf. M. & Köhlin, G. (2009). The Role of Production Risk in

Sustainable Land-Management Technology Adoption in the EthiopianHighlands. Gothenburg University Working Papers in Economics, 407.

Khandker, S. R,. Koolwal, G. B. & Samad, H. A. (2010). Handbook on Impact

Evaluation: Quantitative Methods and Practices. Washington D. C.. TheWorld Bank.

Labeaga, J. M. (1999). A Double-Hurdle Rational Addiction Model withHeterogeneity: Estimating the Demand for Tobacco. Journal of

Econometrics 93, 49-72.Lin, C.T. J. & Milon, J. W. (1993). Attribute and Safety Perceptions in a Double-

Hurdle Model of Shellfish Consumption. American Journal of Agricultural

Economics. 75, 724-729.Manning, W. G., Basu, A. & Mullahy, J. (2002). Modeling Costs with Generalized

Gamma Regression.

Manning, W. G. & Mullahy, J. (2001). Estimating log models: to transform or not totransform? Journal of Health Economics 20.

Ornelas Almaraz G. A. (2012). Patterns of Tobacco Consumption in Mexico –Current Perspective. PhD thesis. The University of East Anglia.

Oseni, M. O. (2015). Assessing the Consumers’ Willingness to Adopt a PrepaymentMetering System In Nigeria. Energy Policy. 86. 154-165.

OXFAM. (2012). Country profile of Ethiopia, Canada.Pycroft, J. (2008). The Adoption and Productivity of Modern Agricultural

Technologies in the Ethiopian Highlands: A cross-sectional analysis ofmaize production in the West Gojam zone. Prepared for presentation at the

Centre for the Study of African Economies (CSAE) Conference. Economic

Development in Africa. 15th-17th March 2008. University of Sussex.Rosenbaum, P. R. & Rubin, D. B. (1985). Constructing a Control Group using

Multivariate Matched Sampling Methods that Incorporate the PropensityScore. The American Statistician. 39, 33-38.

Salmon, C. & Tanguy, J. (2015). Rural Electrification and Household Labor Supply:Evidence from Nigeria. Institute for Labor Studies and Public Policies.


146

Saz-Salazar, S. D. and Rausell-Koster, P. (2008). A Double-Hurdle Model of UrbanGreen Areas Valuation: Dealing with Zero Responses. Landscape and

Urban Planning 84. 241-251.Steiner, P. M. & Cook, D. 2013. Matching and Propensity Scores. The Oxford

Handbook of Quantitative Methods. New York: Oxford University Press.Tauras, J. A. (2005). An Empirical Analysis of Adult Cigarette Demand. Eastern

Economic Journal. 31Terefe, A. T., Ahmed. M. H. & Gebremariam, G. G. (2013). Adoption and Extent of

use of Organic Fertilizer in Arsi Negelle District. Oromia Regional State ofEthiopia: What are the Sources? Advanced Journal of Agricultural

Research. 1. 061-071.UNDP. (2013). Human Development Report 2013. The Rise of the South: Human

Progress in a Diverse World. New York.Yen, S. T. & Jensen, H. H. (1996). Determinants of Household Expenditures on

Alcohol. The Journal of Consumer Affairs. 30. 48-67.Yesuf, M. & Köhlin, G. (2009). Market Imperfections and Farm Technology

Adoption Decisions: Case Study from the Highlands of Ethiopia.Department of Economics Göteborg University.

Yu, B., Nin-Pratt, A., Funes, J. & Gemessa, S. A. (2011). Cereal Production andTechnology Adoption in Ethiopia. The Ethiopia Strategy Support Program

II (ESSP II) Working Paper, 31.


147

AppendicesAppendix 1

VariableChemical fertilizer adoption

No Yes Total T-valueHousehold Characteristics

Age 50.13 46.85 48.68 5.2

Household size 5.16 5.20 5.18 -0.51

Education .43 .82 .69 -4.78Number of oxen 1.13 1.52 1.38 -8.6

Livestock (TLU) 3.50 3.9 3.8 -3.1Market Access and technology diffusion

Distance from woreda town (km) 26.54 18.41 21.22 11.22Distance from the nearest asphalt(km) 37.00 34.51 35.36 1.34

Distance from the nearest weeklymarket 13.2 8.7 9.78 8.70

Distance from major urban 92.40 70.60 78.10 7.37Plot Characteristics

Distance from homestead 1.00 .833 .926 1.71Plot Slope (percent) 20.33 11.37 14.47 16.12

Plot potential wetness index 12.24 12.95 12.7 -6.80

Plot size (hectare) .137 .198 .164 -9.13

Plot elevation 2085 2206.6 2164 -5.97

Average crop yield in kilogram per hectareCrop yield (total sample. KG) 1393.6 1495.58 1461 -0.71Crop yield (Maize) 1990.9 1887.6 1921.70 0.4Crop yield (Teff) 971.8 1153.75 1092.76 -0.6Crop yield (Wheat) 1160.97 1423.9 1329.79 -1.8


148

For categorical variables (percent)

VariableAdopt

Total 2 (P-value)No Yes

Extension 13.67 75.71 41.26 867.9(0.000)

Manure 78.49 16.53 38.56 568.1(0.000)

Compost 42.39 12.65 23.28 173.35(0.000)

Certified 75.9 80.2 77.79 5.21(0.023)

Soil quality

Good 31 38 34

14.1 ( 0.001)Fair 47 44.7 46. 06

Poor 21.73 17.21 19.72

Sex(male=1) 87.91 89.37 88.55 1.1405(0.286)

Literate 35.92 38.95 37.26 2.1439(0.143)

Credit 20.39 50.00 33.53 215.80(0.000)

Advisory service 74.52 95.64 83.89 181.08(0.000)

Offarm Employment 7.1 4.53 5.67 6.7653(0.009)

Appendix 2What is the cause of damage of [Crop]?

Freq. Percent Cum.

Too Much Rain |Too Little Rain |Insects |Crop Disease |Weeds |Hail |Frost |Floods |Wild Animals |Birds |Depletion of Soil |Bad Seeds |Other Specify |

4077753051471

2291

301032

9.4118.1217.65

7.0612.0011.06

0.245.182.120.247.062.357.53

9.4127.5345.1852.2464.2475.2975.5380.7182.8283.0690.1292.47

100.00

Total | 425 100.00


149

Appendix 3: Combined average Marginal effects from the double hurdlemodel (for the total sample)

Variable dy/dx Std. Err. z P>|z|

sex* 9.944333 4.77832 2.08 0.037

age -.5473838 .15157 -3.61 0.000

educat~n 5.493141 1.1088 4.95 0.000

agroec~y* 5.974511 6.95786 0.86 0.391

area_h~e -203.391 30.456 -6.68 0.000

dist_r~d -.4154768 .14552 -2.86 0.004

dist_m~t -.1405554 .06623 -2.12 0.034

dist_a~r .0389304 .03852 1.01 0.312

offfarm* -10.25094 5.97141 -1.72 0.086

extens~n* 67.35615 10.433 6.46 0.000

soil_q~y -3.312054 1.50679 -2.20 0.028

dist_h~d -.5872526 .64412 -0.91 0.362

credit* 7.481162 4.12205 1.81 0.070

manure* -20.58218 4.23779 -4.86 0.000

compost* -12.36074 3.37167 -3.67 0.000

plot_s~p -.4574314 .18483 -2.47 0.013

plot_twi 1.901591 .98503 1.93 0.054

plot_s~m .033503 .00577 5.81 0.000

hh_size -2.73941 .90317 -3.03 0.002

livest~u -4.337814 .99638 -4.35 0.000

(*) dy/dx is for discrete change of dummy variable from 0 to 1


150

Appendix 4: Average Marginal Effects after Probit (the first hurdle)| Delta-method

| dy/dx Std. Err. z P>|z|[95% Conf.

Interval]sex | -.0151799 .0245047 -0.62 0.536 -.0632082 .0328484

age | -. -.0017325 .000548 -3.16 0.002 -.0028066 0006584

education | .0069655 .004698 1.48 0.138 -.0022425 .0161735

land_area | .1537459 .0511316 3.01 0.003 .0535298 .2539621

offfarm | .1222236 .0342035 3.57 0.000 .0551861 .1892611

dist_road | -.001349 .0006016 -2.24 0.025 -.0025281 -.0001699

dist_market | -.0016522 .0003157 -5.23 0.000 -.0022709 -.0010335

dist_admctr | -.0001936 .0001401 -1.38 0.167 -.0004681 .0000809

extension | .3687343 .011474 32.14 0.000 .3462458 .3912229

credit | .063195 .01729 3.66 0.000 .0293073 .0970827

manure | -.1838001 .0188956 -9.73 0.000 -.2208347 -.1467655

compost | -.1064715 .0227273 -4.68 0.000 -.1510161 -.0619269

dist_household | -.0047485 .0051995 -0.91 0.361 -.0149394 .0054424

soilquality_poor| -.0616316 .021049 -2.93 0.003 -.1028869 -.0203763

plot_srtmslp | -.0054123 .0008594 -6.30 0.000 -.0070968 -.0037278

plot_twi | -.0048808 .0038969 -1.25 0.210 -.0125185 .0027569

plot_srtm | .0001111 .0000205 5.41 0.000 .0000708 .0001513

agroecology | -.0659415 .0243254 -2.71 0.007 -.1136184 -.0182647

Appendix 5

Poverty Average land holding in hectare Std. err. t-value

Non-poor .17701466 .0045877 3.9394

poor .14111327 .0064037

Total .1687934 .0038446

Difference .0359014 .0091134


151

Appendix 6

What is the reason for not using chemical fertilizers? Freq. Percent Cum.

Ignorance 57 8.74 8.74High Price 84 12.88 21.63Lack of Money 267 40.95 62.58

Non-Availability of Supply 17 2.61 65.18Skeptical of the Outcome 98 15.03 80.37Other Specify 128 19.63 100.00

Total 652 100.00

Why did you not get credit Services? Freq. Percent Cum

Non-Availability of the Service 14 1.14 1.14

Unable to Pay the Loan 597 48.50 49.63

Inadequate Service Provided 164 13.32 62.96

Ignorance 27 2.19 65.15

Does Not Yield Any Results 102 8.29 73.44

Other Specify 327 26.56 100.00

Total 1.231 100.00

Appendix 7Do you use any manure on[Field]?

Summary of livestock tlut-value(diff=0)

Mean Std. Dev.

No 3.7212931 2.4692197 -1.8406

Yes 3.9677777 2.5839421

Total 3.784363 2.5006767


152

0.1

.2.3

.4FG

T(z,

alpha

= 1)

0 1512.4 3024.8 4537.2 6049.6 7562Poverty line (z)

0 1

FGT Curves (alpha=1)

0.2

.4.6

FGT(z

, alph

a = 0)

0 1512.4 3024.8 4537.2 6049.6 7562Poverty line (z)

0 1


Appendix 8


153

Appendix 9Propensity score matching: logistic regressionLogistic regression Number of obs = 540

LR chi2(8) = 118.65Prob> chi2 = 0.0000Log likelihood = -310.16063 Pseudo R2 = 0.1606

chemical_f~r Coef. z Std. Err. P>|z| [95% Conf. Interval]

sex |

age |

education |

livestock_~u |

area_hectar |

credit |

compost |

manure |

_cons |

.1003654

-.0108202

.1993076

-.0408566

1.067947

1.521178

.4280458

.0458289

-.3791707

.2817922

.006532

.0660882

.0317039

.2348623

.2361859

.2353659

.2005118

.4374617

0.36

-1.66

3.02

-1.29

4.55

6.44

1.82

0.23

-0.87

0.722

0.098

0.003

0.198

0.000

0.000

0.069

0.819

0.386

-.4519371

-.0236227

.0697771

-.102995

.6076256

1.058262

-.0332629

-.3471669

-1.23658

.6526679

.0019822

.3288382

.0212819

1.528269

1.984094

.8893546

.4388248

.4782385

0.0

5.1

.15

.2.2

5FG

T(z,

alpha

= 2

)

0 1512.4 3024.8 4537.2 6049.6 7562Poverty line (z)

0 1



154

Sensitivity Analysisrboundslnpcce. Gamma (1 (0.05) 2)Rosenbaum bounds for lnpcce (N = 595 matched pairs)Gamma sig+ sig- t-hat+ t-hat- CI+ CI-11.051.11.151.21.251.31.351.41.451.51.551.61.651.71.751.81.851.91.952

000000000000000000000

000000000000000000000

7.086437.067137.048537.030287.013066.996246.979696.963986.947766.932856.918066.903756.889956.877116.863676.851266.839056.826586.815366.804126.79309

7.086437.105847.124117.141367.158187.173467.187457.201297.213847.225797.237437.248747.259927.270067.280747.290397.300117.309337.317827.325717.33376

7.011456.9914

6.971236.9519

6.933236.915126.897626.881466.864766.849036.834036.819246.804996.791326.777956.765426.752956.740736.729486.717396.70646

7.159257.17745

7.1957.210527.22543

7.23987.253597.267017.279957.291937.303827.314877.325037.33522

7.34527.35447

7.36337.372067.381127.38967

7.3977

* gamma - log odds of differential assignment due to unobserved factors

sig+ - upper bound significance level

sig- - lower bound significance level

t-hat+ - upper bound Hodges-Lehmann point estimate

t-hat- - lower bound Hodges-Lehmann point estimate

CI+ - upper bound confidence interval (a= .95)

CI- - lower bound confidence interval (a= .95)


155

-20 0 20 40 60 80Standardized % bias across covariates

age

livestock_tlu

manure

sex

compost

education

area_hectar

credit

UnmatchedMatched

0 .2 .4 .6 .8 1Propensity Score

Untreated: Off support Untreated: On supportTreated: On support Treated: Off support


156

agricultural technology adoption and rural poverty: a study on smallholders … · 2017-07-27 ·...

Documents